Research map / Large language models / Instruction and fine-tuning

SAGE: Scalable Automated Robustness Augmentation for LLM Knowledge Evaluation

NeurIPS 2026 · Large language models › Instruction and fine-tuning · approach: Efficient LLM fine-tuning systems · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SAGE, a framework that uses a rubric-based verifier (VariantQual) and an SFT+GRPO-trained small generator (VariantGen), both Qwen3-0.6B, to automatically build robustness-augmented multiple-choice benchmarks at far lower cost than human-annotated HellaSwag-Pro.

Problem

Robustness augmentation of knowledge evaluation benchmarks via LLM generate-then-verify pipelines is costly, with low variant yield and unreliable, labor-intensive verification.

Approach

Replace prompting strong LLMs with small fine-tuned models: a rubric verifier (type compliance, label correctness, answer uniqueness) trained on human-labeled seed data, used as the reward model for GRPO optimization of a variant generator initialized by SFT.

Key results

• VariantGen pass rate rises from 90.53% (SFT) to 91.80% with GRPO; SAGE exceeds Qwen2.5-Max on all seven variant types.
• IRA verifier reaches 90.21% ACC and 79.10% AUC vs 87.90% and 64.60% for ERA.
• Ranking of 12 models on SAGE variants correlates with HellaSwag-Pro reference (Spearman rho 0.857 for ARA, 0.964 for CRA); human check finds 86.6% of 350 variants valid; cost ~284 USD vs ~7,064 USD for HellaSwag-Pro (~25x reduction).

Surprising findings

• GRPO optimization of VariantQual gave marginal, inconsistent changes (ACC 90.21% to 90.12%), suspected reward hacking, so the SFT verifier was kept.
• Qwen3-0.6B matched or beat larger Qwen3-1.7B and 4B backbones for both VariantQual and VariantGen.

Limitations

• Evaluated only on multiple-choice benchmarks (HellaSwag primary, MMLU as case study) with seven predefined variant types.
• Generator may exploit verifier blind spots; Scenario Refinement and Critical Testing have lower strict compliance rates.
• Robustness scores are diagnostic and may be affected by variant-induced difficulty changes or artifacts.

Experimental setup

Datasets: HellaSwag HellaSwag-Pro MMLU

Benchmarks: HellaSwag MMLU

Metrics: ACC AUC Recall F1 pass rate OA ARA RLA CRA Spearman rho

Baselines: Qwen2.5-Max prompt-based generation HellaSwag-Pro

Builds on: GRPO HellaSwag-Pro LoRA Bloom's taxonomy

Closest papers by meaning

• Brittlebench: Quantifying LLM robustness via prompt sensitivity (NeurIPS 2026 · similarity 0.88)
• PertEval: Unveiling Real Knowledge Capacity of LLMs with Knowledge-Invariant Perturbations (NeurIPS 2024 · similarity 0.88)
• metabench - A Sparse Benchmark of Reasoning and Knowledge in Large Language Models (ICLR 2025 · similarity 0.88)
• Autonomous Evaluation of LLMs for Truth Maintenance and Reasoning Tasks (ICLR 2025 · similarity 0.88)
• Remarkable Robustness of LLMs: Stages of Inference? (NeurIPS 2025 · similarity 0.88)
• DARG: Dynamic Evaluation of Large Language Models via Adaptive Reasoning Graph (NeurIPS 2024 · similarity 0.87)